fix(serve): a non-object tool_choice.function answered 500 "engine failed"

[kevin9327]

What happens

POST /v1/chat/completions
{"model": "...", "messages": [...],
 "tools": [{"type": "function", "function": {"name": "search"}}],
 "tool_choice": {"type": "function", "function": "search"}}

HTTP/1.1 500 Internal Server Error
{"error":{"message":"The colibri engine failed to process the request.",
          "type":"server_error","param":null,"code":"engine_error"}}

[api] 127.0.0.1 - request failed: 'str' object has no attribute 'get'

Writing the name where the object goes is an ordinary client slip, and the
server already has an answer for it -- generation_options() raises
400 "tool_choice function object must include a name." two lines further
down. It never runs, because the read above it is

name = (choice.get("function") or {}).get("name") or choice.get("name")

which raises AttributeError as soon as function is not an object. That
reaches do_POST's catch-all, and the request comes back as a 5xx that OpenAI
SDKs retry -- against an engine that was never asked anything.

Reach

The same expression is copied into the five renderers that read the forced
tool, and on /v1/chat/completions the prompt is rendered before
generation_options runs, so those crash first. /v1/completions has no
renderer in front and reached the validator's own copy.

arch	before	after
glm, glm53, kimi, deepseek_v4, deepseek_v41	500 (renderer)	400
qwen38	500 (generation_options)	400
olmoe, inkling, qwen36	400 ("tools unsupported")	400
/v1/completions, any arch	500 (generation_options)	400

/v1/messages is unaffected: anthropic_tools() builds its own tool_choice
and already validates it.

The change

Read the member, then check it -- the same shape as the json_schema fix
(#1587) -- through one small helper, so the six call sites that share the
expression share the guard and the existing 400 is what runs. The legacy
{"type": "function", "name": "search"} spelling still forces the tool, and a
well-formed function object renders byte-for-byte as before; both are asserted.

Verification

$ python -m unittest discover -s tests -p "test_openai_server.py" -t tests -k tool_choice
# before: FAILED (failures=19, errors=4)
#   AttributeError: 'str' object has no attribute 'get'   (generation_options)
#   AssertionError: 500 != 400                            (9 arches x 3 shapes, and /v1/completions)
# after:  Ran 6 tests ... OK

Full module and neighbours, on Windows (Python 3.11):

tests/test_openai_server.py        173 OK (baseline) -> 176 OK
tests/test_anthropic_messages.py    28 OK
tests/test_cli_output.py            34 OK
tests/test_family_registry.py       58 OK
tests/test_datapoint.py             20 OK
tests/test_chat_images.py            3 OK

The HTTPTest cases drive a real APIServer over a socket, so what is asserted
is the status the client sees.

🤖 Generated with Claude Code